Rotating speed-temperature-wear shield cutter intelligent sensing method

Abstract

The invention relates to the technical field of drilling of soil layers or rocks, and particularly discloses a rotating speed-temperature-abrasion shield cutter intelligent sensing method. Comprising the steps of micro-synchronous acquisition, coupling feature extraction, physical modeling and synthesis, multi-scale thermal anomaly separation, physical constraint neural network prediction, time sequence evolution prediction, dynamic threshold risk decision, adaptive sampling scheduling and online increment calibration. According to the method, the early warning threshold value is dynamically adjusted based on prediction error distribution, flexible switching among different reliability requirements can be achieved, and the problem of early warning or missing report is avoided. Meanwhile, the sensing sampling frequency and the early warning level are linked, the data resolution can be automatically improved in the high-risk stage, calculation and transmission resources are saved in the low-risk stage, and efficient utilization of the computing power and the bandwidth by the monitoring system is achieved; therefore, distribution drifting caused by geological condition changes and cutter updating can be coped, the effective service life of the model can be prolonged, and shutdown maintenance is almost not needed.

Description

Technical Field

[0001] The present invention belongs to the technical field of drilling in soil layers or rocks, and particularly relates to an intelligent perception method for shield cutters of rotational speed - temperature - wear. Background Technique

[0002] In the field of drilling in soil layers or rocks, as a key component that directly contacts the geotechnical medium and realizes the tunneling function, the wear state of shield cutters directly affects the efficiency, quality, and cost of shield construction. Accurately and real - time perceiving the wear condition of shield cutters is of great significance for timely replacing cutters, avoiding construction failures caused by excessive cutter wear, and optimizing shield tunneling parameters.

[0003] At present, certain progress has been made in the perception technology for shield cutter wear, but there are still many deficiencies.

[0004] Firstly, in terms of heat flux density calculation, the existing technology fails to fully consider the correlation between the tool surface topography and the friction - heat transfer process. The 3D topography of the tool surface has an important impact on the friction heat flux distribution, and traditional methods lack an effective means of combining the 3D topography scan data of the tool surface with the friction - heat transfer equation, resulting in inaccurate calculation of the instantaneous friction heat flux density, and further affecting the accurate assessment of the tool wear condition.

[0005] Secondly, in terms of separating abnormal signals induced by wear, when the existing methods process the heat flux distribution data, it is difficult to effectively separate the high - frequency heat anomaly components directly related to local tool defects. The abnormal signals generated during the tool wear process are often masked by low - frequency environmental disturbances (such as overall temperature drift) and structural noises (such as mechanical vibrations), resulting in a large amount of noise components in the subsequent neural network input data, reducing the signal - to - noise ratio and accuracy of the overall monitoring, and making the judgment of tool wear prone to deviation.

[0006] Finally, in the construction of the wear rate prediction model, the existing models have deficiencies in capturing the long - term and short - term dependence relationships of the wear rate. The wear rate of shield cutters is affected by both instantaneous sudden factors and the long - term cumulative wear process. However, traditional models are difficult to take both aspects into account simultaneously, resulting in poor accuracy of the prediction results in terms of short - term response and long - term planning, and unable to provide a reliable decision - making basis for shield construction.

[0007] In response to this, the present application proposes an intelligent perception method for shield cutters of rotational speed - temperature - wear to solve the above problems. Summary of the Invention

[0008] The purpose of the present invention is to provide an intelligent perception method for shield cutters of rotational speed - temperature - wear to solve the problems raised in the above background technique.

[0009] To achieve the above object, the present invention provides the following technical solutions: An intelligent perception method for shield cutter with rotational speed-temperature-wear, comprising: Arranging a thermocouple array and an optical encoder on the cutter surface to synchronously collect temperature time series signals and rotational speed time series signals, and constructing a preliminary correlation matrix; performing mutual information analysis on the correlation matrix to extract temperature-rotational speed coupling features and obtain a feature set; Least square fitting the feature set and the 3D surface topography scan data of the cutter surface through a friction-heat transfer equation to calculate the instantaneous frictional heat flux density and obtain a heat flux distribution model; performing wavelet packet decomposition on the heat flux distribution model to separate the high-frequency heat anomaly components induced by wear and obtain anomaly components; Inputting the feature set and the anomaly components into a physically constrained neural network model to predict the instantaneous wear rate and obtain an initial wear rate; Time series evolution and online calibration step: Inputting the initial wear rate and the latest temperature-rotational speed historical sequence into an LSTM-Transformer model to predict the future wear process, and performing online incremental calibration in combination with the measured wear value after actual replacement to obtain a final optimized prediction result; Using the prediction result and the latest collected temperature-rotational speed historical sequence as the input of the LSTM-Transformer model to predict the future wear process and obtain a prediction sequence; performing dynamic threshold determination based on the P-value control on the prediction sequence to generate a cutter risk warning state; Using the warning state to adjust the sampling frequency and alarm level of the thermocouple and the encoder to obtain an optimized acquisition strategy; comparing the measured wear value of the cutter after actual replacement with the prediction sequence to update the weights of the neural network and the time series model online to obtain an optimized model.

[0010] Preferably, the performing mutual information analysis on the correlation matrix to extract temperature-rotational speed coupling features and obtain a feature set includes: calculating the mutual information between temperature and rotational speed through a mutual information evaluation method to identify the most effective coupling quantity for wear prediction from the correlation matrix, and selecting several features with the largest amount of information and the lowest redundancy based on the maximum correlation minimum redundancy criterion; The formula of the mutual information evaluation method is: ; where, T: random variable of cutter surface temperature, value set T = {t1,..., tM}; ω: random variable of cutter rotational speed, value set Ω = {ω1,..., ωN}; p(t, ω): joint probability density, representing the observation frequency when the temperature is t and the rotational speed is ω; p(t), p(ω): Marginal probability density, which are the observation frequencies of temperature or rotational speed respectively.

[0011] Preferably, in the step of calculating the instantaneous frictional heat flux density by least-squares fitting of the feature set and the 3D topography scanning data of the tool surface through the friction-heat transfer equation to obtain the friction-heat transfer equation in the heat flux distribution model, the formula is: ; where q is the instantaneous frictional heat flux density, that is, the frictional heat flux generated per unit area, with the unit of W / m²; h is the surface heat transfer coefficient, reflecting the conduction and convection capabilities between the tool and the environment, with the unit of W / (m²·K); Ts is the real-time temperature of the tool surface, measured by a thermocouple array, with the unit of °C; Ta is the ambient temperature, which can be the temperature of surrounding rock cuttings or linings, with the unit of °C; μ is the friction coefficient, which is related to the tool material and rock type; p is the tool-rock contact pressure, obtained by feedback from a force sensor or a robotic arm, with the unit of Pa; v is the tool linear velocity, equal to the rotational speed ω multiplied by the radius, with the unit of m / s.

[0012] Preferably, in the step of inputting the feature set and the abnormal component into the physics-informed neural network model to predict the instantaneous wear rate and obtain the initial wear rate, it includes adding constraints to the loss function and adopting the physics-informed loss function to ensure that the data-driven model does not violate the friction-heat transfer mechanism; The formula of the physics-informed loss function is: ; where N is the total number of training samples; , yi are the wear rates predicted by the neural network and actually measured for the i-th sample, with the unit of mm / h; ΔEi pred : The thermal energy change deduced from the predicted wear amount for the i-th sample, with the unit of J; ΔEi phys : The physical thermal energy change calculated according to the friction-heat transfer equation for the i-th sample, with the unit of J; λ is the weight of the constraint term, used to balance the fitting accuracy and physical consistency (typical value 0.1–1.0).

[0013] Preferably, in the step of performing wavelet packet decomposition on the heat flux distribution model to separate the high-frequency thermal anomaly components induced by wear and obtain the abnormal component, it includes: using the db4 wavelet basis to decompose the heat flux distribution model to level 4 to extract the high-frequency thermal anomaly components in the sub-bands of level 3 and level 4.

[0014] Preferably, the LSTM-Transformer model includes a bidirectional LSTM encoder and a multi-head self-attention decoder, where the number of attention heads is not less than 8, which is used to capture the long-term and short-term dependencies of the historical sequence; Online incremental calibration adopts a sample replay strategy, where the single weight update ratio does not exceed 5%.

[0015] Preferably, the feature set and the abnormal component are input into the physical constraint neural network model to predict the instantaneous wear rate, and in obtaining the initial wear rate, the neural network structure is a three-layer fully connected network, the number of input layer nodes is equal to the sum of the dimensions of the feature set and the abnormal component, the middle layer includes an energy conservation constraint term, and the output layer is a single-node prediction of the instantaneous wear rate.

[0016] Preferably, in the dynamic threshold determination based on the P-value control for the prediction sequence to generate the tool risk warning state, the P-value control calculates the distribution parameter of the prediction error based on the sliding window method and dynamically adjusts the warning threshold to the 95% confidence level.

[0017] Preferably, in using the warning state to adjust the sampling frequencies and alarm levels of the thermocouple and the encoder to obtain the optimized acquisition strategy, the sampling frequencies are divided into three levels according to the warning state: high risk of ≥20 kHz, medium risk of 10 - 20 kHz, and low risk of 5 - 10 kHz for execution.

[0018] Preferably, in comparing the measured value of the actual tool wear after replacement with the prediction sequence to update the weights of the neural network and the time series model online to obtain the optimized model, the model update adopts an incremental learning strategy based on sample replay, and only 5% of the weights are updated each time for calibration.

[0019] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention extracts the coupling features between temperature and rotational speed through mutual information analysis. The method can capture the implicit wear signals that are difficult to reflect by a single sensing dimension, significantly improving the sensitivity to the micro wear stage of the tool. Compared with the method that simply relies on temperature or vibration, the multi-source coupling features can reveal the wear trend at an early stage, thus giving an early warning and reducing the "blind area" time; integrating the friction - heat transfer physical equation into the heat flux density fitting and the neural network loss function not only closely combines the empirical data-driven and theoretical mechanisms, but also suppresses overfitting and physical violation phenomena through the energy conservation constraint in model training. This not only enhances the robustness of the model under unknown working conditions, but also improves the interpretability of the results, facilitating on-site engineers to conduct fault diagnosis based on physical mechanisms.

[0020] (2) The present invention adopts wavelet packet multi-scale decomposition to effectively separate the high-frequency thermal anomalies directly associated with local tool defects, enabling the subsequent neural network input to focus on the most representative abnormal components. This processing has significant advantages in suppressing low-frequency environmental disturbances (such as overall temperature drift) and structural noises (such as mechanical vibrations), thereby improving the signal-to-noise ratio and accuracy of overall monitoring. The prediction model is designed as a hybrid structure of LSTM and multi-head self-attention, which can simultaneously capture the short-term fluctuations and long-term trends of the wear rate. This kind of time series modeling method can not only reflect instantaneous sudden wear but also take into account the cumulative attenuation law, making the prediction results highly accurate in both short-term response and long-term planning.

[0021] (3) The present invention dynamically adjusts the warning threshold based on the prediction error distribution, which can flexibly switch between different reliability requirements to avoid the problems of "premature warning" or "false alarm". At the same time, by linking the sensing sampling frequency with the warning level, it can automatically improve the data resolution in the high-risk stage and save computing and transmission resources in the low-risk stage, realizing the efficient utilization of computing power and bandwidth by the monitoring system. Through incremental learning by comparing the measured wear after actual replacement with the predicted value, this solution can continuously self-calibrate during operation and correct model biases in a timely manner. This mechanism can not only cope with the distribution drift caused by geological condition changes and tool updates but also extend the effective life of the model, with almost no need for shutdown maintenance. Brief Description of the Drawings

[0022] Figure 1 It is a flowchart of the intelligent perception method for shield cutter with rotational speed-temperature-wear of the present invention. Detailed Embodiments

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] Embodiment 1: Please refer to Figure 1 As shown, the intelligent perception method for shield cutter with rotational speed-temperature-wear includes: S1: Arrange a thermocouple array and an optical encoder on the tool surface to synchronously collect the temperature time series signal and the rotational speed time series signal, and construct a preliminary correlation matrix; perform mutual information analysis on the correlation matrix to extract the temperature-rotational speed coupling characteristics and obtain a feature set; The mutual information analysis of the correlation matrix described in step S1 is used to extract the temperature-rotation coupling characteristics, and the obtained feature set includes: calculating the mutual information between temperature and rotation speed through the mutual information evaluation method to identify the most effective coupling quantity for wear prediction from the correlation matrix, and selecting several features with the largest amount of information and the lowest redundancy based on the maximum correlation and minimum redundancy criterion; The formula of the mutual information evaluation method is: ; where T: the random variable of the tool surface temperature, and the value set T = {t1, …, tM}; ω: the random variable of the tool rotation speed, and the value set Ω = {ω1, …, ωN}; p(t, ω): the joint probability density, representing the observation frequency when the temperature is t and the rotation speed is ω; p(t), p(ω): the marginal probability densities, which are the observation frequencies of temperature or rotation speed respectively; By calculating I(T; ω), the temperature-rotation coupling characteristics that are most relevant to the temperature rise change and have the least redundancy can be screened out to construct a feature set, significantly improving the sensitivity and generalization ability of the subsequent model to the working condition switching; S2: Perform the least squares fitting of the feature set and the tool surface 3D topography scan data through the friction-heat transfer equation to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model; perform wavelet packet decomposition on the heat flux distribution model to separate the high-frequency heat anomaly components induced by wear and obtain the abnormal components; The formula of the friction-heat transfer equation in the process of performing the least squares fitting of the feature set and the tool surface 3D topography scan data to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model is: ;

[0025] where q: the instantaneous friction heat flux density, that is, the friction heat flux generated per unit area, with the unit of W / m²; h: the surface heat transfer coefficient, reflecting the conduction and convection capabilities between the tool and the environment, with the unit of W / (m²·K); Ts: the real-time temperature of the tool surface, measured by the thermocouple array, with the unit of °C; Ta: the ambient temperature, which can be the surrounding rock debris or lining temperature, with the unit of °C; μ: the friction coefficient, related to the tool material and rock type; p: the tool-rock contact pressure, obtained by the feedback of the force sensor or the robotic arm, with the unit of Pa; v: the tool linear velocity, equal to the rotation speed ω multiplied by the radius, with the unit of m / s; Using this equation, the coupling characteristics can be converted into a heat flux distribution model to achieve the physical quantification of the tool force and thermal field. The temperature - rotational speed characteristics are mapped into a heat flux distribution model to capture the heat source changes during the tool friction process, providing an accurate physical basis for high - frequency thermal anomaly separation and wear prediction; S3: Input the feature set and the abnormal component into the physics - informed neural network model for predicting the instantaneous wear rate to obtain the initial wear rate; The step of inputting the feature set and the abnormal component into the physics - informed neural network model for predicting the instantaneous wear rate to obtain the initial wear rate includes adding constraints to the loss function and using a physics - informed loss function to ensure that the data - driven model does not violate the friction - heat transfer mechanism; The formula for the physics - informed loss function is: ; where, N: the total number of training samples; , yi: the wear rate predicted by the neural network and the actual measured wear rate of the i - th sample, with the unit of mm / h; ΔEi pred : the change in thermal energy deduced from the predicted wear amount of the i - th sample, with the unit of J; ΔEi phys : the physical change in thermal energy calculated according to the friction - heat transfer equation for the i - th sample, with the unit of J; λ: the weight of the constraint term, used to balance the fitting accuracy and physical consistency (typical value 0.1 - 1.0).

[0026] The first - term LMSE ensures the fitting accuracy of the predicted wear rate, and the second - term Lcons enforces the output to satisfy the energy conservation, significantly improving the physical interpretability of the model and its robustness under unknown working conditions; S4: Time - series evolution and online calibration step: Input the initial wear rate and the latest temperature - rotational speed historical sequence into the LSTM - Transformer model for predicting the future wear process, and perform online incremental calibration in combination with the actual measured wear value after replacement to obtain the final optimized prediction result.

[0027] S5: Use the prediction result and the latest collected temperature - rotational speed historical sequence as the input of the LSTM - Transformer model to predict the future wear process and obtain a prediction sequence; perform dynamic threshold determination based on the P - value control on the prediction sequence to generate the tool risk warning state; S6: Use the warning state to adjust the sampling frequency and alarm level of the thermocouple and encoder to obtain an optimized acquisition strategy; compare the actual measured wear value of the tool after replacement with the prediction sequence to update the weights of the neural network and the time - series model online to obtain an optimized model.

[0028] Specifically, the wavelet packet decomposition of the heat flux distribution model is used to separate the high-frequency heat anomaly components induced by wear. The abnormal components obtained include: using the db4 wavelet basis, decomposing the heat flux distribution model to level 4 to extract the high-frequency heat anomaly components in the third and fourth level sub-bands.

[0029] Specifically, the LSTM-Transformer model includes a bidirectional LSTM encoder and a multi-head self-attention decoder, where the number of attention heads is not less than 8, which is used to capture the long-term and short-term dependencies of the historical sequence; Online incremental calibration adopts a sample replay strategy, where the single weight update ratio does not exceed 5%.

[0030] Specifically, inputting the feature set and the abnormal components into the physically constrained neural network model to predict the instantaneous wear rate. In the initial wear rate obtained, the neural network structure is a three-layer fully connected network. The number of nodes in the input layer is equal to the sum of the dimensions of the feature set and the abnormal components. The middle layer includes an energy conservation constraint term, and the output layer is a single-node prediction of the instantaneous wear rate.

[0031] Specifically, performing dynamic threshold determination based on the P-value control on the prediction sequence to generate the tool risk warning state. In the P-value control, the distribution parameters of the prediction error are calculated based on the sliding window method, and the warning threshold is dynamically adjusted to the 95% confidence level.

[0032] Specifically, using the warning state to adjust the sampling frequencies and alarm levels of the thermocouple and the encoder. In the optimized acquisition strategy obtained, the sampling frequencies are divided into three levels according to the warning state: high risk of ≥20 kHz, medium risk of 10 - 20 kHz, and low risk of 5 - 10 kHz for execution.

[0033] Specifically, comparing the measured value of the tool wear after actual replacement with the prediction sequence to update the weights of the neural network and the time series model online. In the optimized model obtained, the model update adopts an incremental learning strategy based on sample replay, and only 5% of the weights are updated each time for calibration.

[0034] As can be seen from the above, by extracting the coupling features between temperature and rotational speed through mutual information analysis, the method can capture the hidden wear signals that are difficult to reflect by a single sensing dimension, and significantly improve the sensitivity to the micro wear stage of the tool. Compared with the method that simply relies on temperature or vibration, the multi-source coupling features can reveal the wear trend at an early stage, so as to give early warnings and reduce the "blind area" time; Integrate the friction - heat transfer physical equation into the heat flux density fitting and the neural network loss function, which not only closely combines empirical data - driven with theoretical mechanisms, but also suppresses overfitting and physical violation phenomena through energy conservation constraints during model training. This not only enhances the robustness of the model under unknown working conditions, but also improves the interpretability of the results, facilitating on - site engineers to conduct fault diagnosis based on physical mechanisms; Adopt wavelet packet multi - scale decomposition to effectively separate the high - frequency thermal anomalies directly associated with local tool defects, enabling the subsequent neural network input to focus on the most representative abnormal components. This processing has significant advantages in suppressing low - frequency environmental disturbances (such as overall temperature drift) and structural noises (such as mechanical vibrations), thereby improving the signal - to - noise ratio and accuracy of overall monitoring.

[0035] Example 2: Tool wear monitoring under soft soil conditions: 1. Working conditions and objectives Under the conditions of soft soil layers with high moisture content (average moisture content of 25%), soil cohesionless, soil standard penetration resistance N value of about 2 - 4, and hardness of about 0.5 MPa, focus on verifying the accurate prediction of the real - time wear rate of this solution. The objective is to control the prediction error within 5% within 1 hour.

[0036] 2. Data acquisition Temperature data: Use an 8 - point array of Omega brand K - type thermocouples (model TC - 08) with a point spacing of 20 mm, attached to the cutting edge of the tool side, and collect the sequential temperature Ts,i. The data is transmitted to the edge computing unit at a frequency of 1 kHz through an NI USB - 6210 data acquisition card.

[0037] Rotation speed data: Adopt a Renishaw optical encoder (resolution 2000 PPR), installed at the end of the tool holder, and set the sampling frequency to 5 kHz to output the rotation speed ωi in real - time.

[0038] Three - dimensional topography data: Use a Faro Focus 3D laser scanner to perform rotational scanning on the tool surface with a point cloud accuracy of 0.1 mm for constructing a reference geometric model.

[0039] Ambient temperature: Arrange an SHT35 temperature - humidity sensor near the tool, with a sampling frequency of 1 Hz, for recording the ambient temperature Ta.

[0040] All sensors are aggregated to an NVIDIA Jetson AGX edge computing platform through industrial Ethernet and stored in the local SSD in real - time.

[0041] 3. Data processing Construct the correlation matrix A Synchronously collect N=10000 {Ts,i, ωi} pairs within 10 seconds to form a matrix A=[(Ts,1,ω1),(Ts,2,ω2),…,(Ts,N,ωN)] Mutual information feature extraction A sliding window with a length of 500 points is used to calculate the mutual information I(T;ω) for each window segment.

[0042] Based on the mRMR algorithm, the 8-dimensional features with the highest mutual information value are selected to form the feature set B={b1,b2,…,b8}.

[0043] Friction-heat transfer least squares fitting Calculate the linear velocity vi=ωi / 60×(1.2 / 2)m / s.

[0044] Using the Fit Function qi=1500 (Ts,i−Ta)+0.4×2×vi The least squares fitting is performed on each sampling point to obtain the heat flow series {qi}.

[0045] Wavelet packet multi-scale decomposition Perform db4, 4-level decomposition on {qi}, retaining the 3rd and 4th level high-frequency coefficients to form abnormal components D={d1,d2,…}.

[0046] Physically Constrained Neural Network Prediction Network structure: input layer dimension 8+2 (main statistics of B and D); two hidden layers (number of nodes 64 and 32); output layer is single-node wear rate prediction.

[0047] The loss function is shown in the following formula: ; An energy conservation constraint term of λ=0.5 is introduced.

[0048] Time series evolution prediction and online calibration: The instant rate sequence { } and the past 1 000 points {Ts,i,ωi} history are written into a bidirectional LSTM + multi-head attention decoder to predict the wear curve for the next 1 hour.

[0049] After each 1 h of monitoring, the cutter teeth were replaced manually and the average wear rate was measured. The network weight was updated by replaying the samples with a value not exceeding 3%.

[0050] Instant wear rate based on neural network After prediction, the rate sequence is input into the trained LSTM-Transformer model together with the historically collected temperature Ts,i and rotational speed ωi acquired synchronously: Input structure: Sequence length: 1000 points (i.e., a 10 s historical window).

[0051] Each point contains 3 dimensions: temperature, rotational speed, and instantaneous wear rate.

[0052] Data normalization is processed using dynamically updated historical means and standard deviations.

[0053] Model structure: LSTM: Bidirectional structure, with 64 single-layer nodes.

[0054] Transformer: Adopts a 4-head attention mechanism, with each head having a dimension of 16.

[0055] Output: A predicted sequence of tool wear rates per minute for the next 60 minutes {f1, f2, …, f60}.

[0056] Dynamic threshold setting method: Based on the model prediction error residual sequence in the past 2 hours, a sliding window statistical distribution is constructed to calculate the wear prediction confidence interval at the current time point.

[0057] Set the confidence level α = 95%, and calculate the dynamic threshold θt using the P-value control function of the prediction error.

[0058] If any point in the current or future prediction sequence exceeds the threshold θt, a risk warning status G is generated and downstream scheduling is triggered.

[0059] For example: In the 38th minute prediction, the predicted value for the 15th minute in the future is 0.052 mm / h, which exceeds the dynamic threshold of 0.048 mm / h (based on the 95% confidence interval in the previous two hours). The system immediately marks this point as "moderate risk" and records the trigger time.

[0060] Sampling scheduling mechanism: According to the warning status generated by the previous prediction result, the system automatically adjusts the sensor sampling strategy as shown in Table 1 below: Table 1

[0061] The system controls the sampling rates of the NI module and the encoder through the CAN bus to ensure that the data granularity during high-risk periods meets the requirements of model tracking.

[0062] Actual measurement feedback and incremental learning: After each manual maintenance or replacement of the cutter head, measure and record the actual wear amount y_actual.

[0063] Retrieve the historical prediction sequence {f_t} corresponding to this time period and calculate the overall error.

[0064] Adopt the sample replay incremental learning mechanism: Sample 100 input-output pairs from the time period with the largest prediction error.

[0065] Fine-tune the existing neural network model using a micro learning rate of 1×10 -5 , and only update the end weights.

[0066] Control the weight update amplitude to be less than 3% of the overall model parameters to prevent "catastrophic forgetting";

[0067] 4. Performance comparison is shown in Table 2 below: Table 2

[0068] As can be seen from the above, under the conditions of soft soil with high water content, the average error of this scheme is reduced by about 80% compared with the traditional single-temperature method, significantly improving the early warning accuracy.

[0069] As can be seen from the above, designing the prediction model as a hybrid structure of LSTM and multi-head self-attention can capture both the short-term fluctuations and long-term trends of the wear rate. This kind of time series modeling method can not only reflect the instantaneous sudden wear but also take into account the cumulative attenuation law, making the prediction results highly accurate in both short-term response and long-term planning; Dynamically adjusting the early warning threshold based on the prediction error distribution can flexibly switch between different reliability requirements, avoiding the problems of "premature early warning" or "false negative". At the same time, linking the sensing sampling frequency with the early warning level can automatically increase the data resolution in the high-risk stage and save computing and transmission resources in the low-risk stage, realizing the efficient utilization of computing power and bandwidth by the monitoring system.

[0070] Example 3: Tool wear monitoring under hard rock formation conditions: 1. Working conditions and objectives Operating under hard sandstone (standard penetration resistance N value is about 50, hardness is 5 MPa), in order to cope with higher frictional resistance and heat load and verify the robustness of this scheme in extremely harsh environments, the target prediction error is <5%.

[0071] 2. Data acquisition Thermocouple array: OMEGA 12-point K-type, point distance 15 mm; sampling frequency 2 kHz, transmitted through Advantech ADAM-4017 module.

[0072] Optoelectronic encoder: Renishaw RGH24 series, 5,000 PPR, sampling frequency 10 kHz, combined with robotic arm rotation angle compensation.

[0073] 3D laser scanning: Leica BLK360, accuracy 0.05 mm, the scanned point cloud is denoised by PCL (Point Cloud Library).

[0074] Environmental acquisition: Use PT100 platinum resistance to measure the environmental temperature Ta in the silo, and use a self-designed soil hardness sensing probe to obtain real-time hardness values.

[0075] 3. Data processing Time synchronization and correlation matrix Align the timestamps of all devices through Precision Time Protocol (PTP).

[0076] Synchronously collect 10,000 groups of data within 5 s to form a correlation matrix A.

[0077] Coupling feature screening Calculate the mutual information with a 200-point sliding window, and select the most sensitive 10-dimensional features to form a feature set.

[0078] Enhanced friction-heat transfer fitting The formula is adjusted to: qi = 1800 (Ts,i − Ta) + 0.6×4×vi + α Hi Where Hi is the geological hardness value, and α = 50 W / (m²·MPa) is the empirical coupling coefficient.

[0079] Simultaneously estimate α by least squares fitting.

[0080] Wavelet packet decomposition and noise suppression: Except for db4, 4-level decomposition, first perform wavelet denoising (threshold hard and soft combination) on {qi}, and then extract high-frequency anomalies.

[0081] Depth prediction and calibration The physical constraint network is the same as in Example 1, but λ is adjusted to 0.7, and the incremental learning weight update rate is 5%.

[0082] LSTM-Transformer is deployed on two layers of 128-unit LSTM, 12 heads of attention, and Dropout 0.2 to prevent overfitting.

[0083] After completing the immediate wear prediction with geological hardness correction, a more complex multivariate prediction model is constructed using a structurally enhanced LSTM-Transformer: Input structure: The historical window has 2000 points, and each point contains 4 dimensions: temperature, rotational speed, wear rate, and rock formation hardness.

[0084] All inputs are standardized, and the hardness is taken as a relative value (normalized to 0-1).

[0085] Model structure: LSTM: Bidirectional two-layer, 128 units.

[0086] Transformer: 12 heads of attention, hidden layer dimension 64.

[0087] Output: A prediction sequence for the next 90 minutes, with 1 point every 5 minutes, and a total of 18 points of wear rate sequence are output.

[0088] Dynamic threshold algorithm: Add a multivariate residual control mechanism: Consider the error covariance of three pairs: temperature-wear, rotational speed-wear, and hardness-wear simultaneously.

[0089] Construct a joint P-value model (multivariate statistics): ; where Pt: The multivariate joint P-value at the current time point t, used to determine whether the current wear prediction significantly deviates from the normal distribution; : The cumulative distribution function (CDF) of the chi-square distribution with k degrees of freedom, used to calculate the position of the observed value in the distribution; k: The dimension of the variable, which is 3 in this embodiment (corresponding to the coupled residual vectors between temperature, rotational speed, hardness, and wear); rt∈Rk: The residual vector at time point t, that is, the deviation combination between the model prediction value and the measured value: Σ∈Rk×k: The residual covariance matrix, reflecting the joint change structure between variables; : The square of the Mahalanobis distance, measuring the degree of deviation of the current residual from the historical distribution; Dynamically adjust the warning level (such as high, medium, low), and automatically suggest whether manual review is required according to the geological conditions; Edge device scheduling: Use the Jetson Xavier module to implement local model execution and policy switching, ensuring that the system can still operate stably in complex and high-vibration environments.

[0090] When the prediction is "high risk", the system automatically increases the thermocouple to 5 kHz, the encoder to 10 kHz, and activates the backup acquisition channel for redundancy.

[0091] Model self-update mechanism: After each round of maintenance, upload the measured wear amount as a label and add it to the "rolling training buffer".

[0092] Adopt online pruning optimization + fine-tuning method for training: Utilize "layered learning rate": Freeze the underlying feature extraction layer and only fine-tune the intermediate GRU units.

[0093] Add Dropout(0.2) to prevent overfitting.

[0094] After each online update is completed, automatically evaluate the error change of the model on the validation set to ensure that the incremental update does not introduce a performance decline;

[0095] 4. Performance comparison is shown in Table 3 below: Table 3

[0096] As can be seen from the above, under the hard rock high-pressure friction working condition, the accuracy of this solution is further improved after adding geological hardness coupling and noise suppression, and the prediction error remains at about 2.5%, meeting the on-site real-time early warning requirements.

[0097] As can be seen from the above, through the comparative tests under the above two working conditions, it can be seen that this solution can maintain extremely high prediction accuracy under different geological conditions, far exceeding the traditional single temperature monitoring technology, fully reflecting the innovative value and engineering applicability of the "rotation speed - temperature - wear" coupling and physical constraint algorithm.

[0098] As can be seen from the above, through incremental learning by comparing the measured wear and predicted values after actual replacement, this solution can continuously self-calibrate during operation and timely correct the model deviation. This mechanism can not only cope with the distribution drift caused by geological condition changes and tool updates, but also extend the effective life of the model, and almost no shutdown maintenance is required.

[0099] This solution collaboratively optimizes from multiple levels such as multi-physical quantity acquisition, physical modeling, signal processing, deep learning, and online calibration, forming a closed-loop perception - prediction - feedback system. This system can maintain high stability and accuracy under complex and changeable strata and working conditions, has good scalability and portability, can be extended to other multi-physical sensing monitoring scenarios (such as bearing wear, motor faults, etc.), and has broad application prospects and economic benefits.

[0100] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0101] In the accompanying drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved. For other structures, reference can be made to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other.

[0102] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent perception method for shield cutters of rotational speed - temperature - wear, characterized in that It includes the following steps: Arrange a thermocouple array and an optical encoder on the tool surface to synchronously collect temperature time series signals and rotational speed time series signals, and construct a preliminary correlation matrix; Perform mutual information analysis on the correlation matrix to extract temperature-rotational speed coupling features and obtain a feature set; Perform least squares fitting on the feature set and the 3D topography scan data of the tool surface through the friction-heat transfer equation to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model; perform wavelet packet decomposition on the heat flux distribution model to separate the high-frequency heat anomaly components induced by wear and obtain anomaly components; Input the feature set and the anomaly components into a physically constrained neural network model to predict the instantaneous wear rate and obtain an initial wear rate; Input the initial wear rate and the latest temperature-rotational speed historical sequence into the LSTM-Transformer model to predict the future wear process, and perform online incremental calibration in combination with the measured value of the actual wear after replacement to obtain a final optimized prediction result; Use the prediction result and the latest collected temperature-rotational speed historical sequence as the input of the LSTM-Transformer model to predict the future wear process and obtain a prediction sequence; perform dynamic threshold determination based on the P-value control on the prediction sequence to generate a tool risk warning state; Use the warning state to adjust the sampling frequency and alarm level of the thermocouple and the encoder to obtain an optimized acquisition strategy; compare the measured value of the actual wear of the tool after replacement with the prediction sequence to update the weights of the neural network and the time series model online to obtain an optimized model.

2. The intelligent perception method for the shield cutter's rotational speed-temperature-wear according to claim 1, characterized in that, The performing mutual information analysis on the correlation matrix to extract temperature-rotational speed coupling features and obtain a feature set includes: calculating the mutual information between temperature and rotational speed through a mutual information evaluation method to identify the most effective coupling quantity for wear prediction from the correlation matrix, and selecting several features with the largest amount of information and the lowest redundancy based on the maximum correlation minimum redundancy criterion; The formula of the mutual information evaluation method is: ； where, T: the random variable of the tool surface temperature, and the value set T = {t1,..., tM}; ω: the random variable of the tool rotational speed, and the value set Ω = {ω1,..., ωN}; p(t, ω): the joint probability density, representing the observation frequency when the temperature is t and the rotational speed is ω; p(t), p(ω): the marginal probability densities, which are the observation frequencies of temperature or rotational speed respectively.

3. The intelligent perception method for shield cutter of rotational speed - temperature - wear according to claim 1, wherein The formula of the friction-heat transfer equation in the performing least squares fitting on the feature set and the 3D topography scan data of the tool surface through the friction-heat transfer equation to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model is: ； where, q: the instantaneous friction heat flux density, that is, the friction heat flux generated per unit area, with the unit of W / m²; h: the surface heat transfer coefficient, reflecting the conduction and convection capabilities between the tool and the environment, with the unit of W / (m²·K); Ts: the real-time temperature of the tool surface, measured by the thermocouple array, with the unit of °C; Ta: the ambient temperature, which can be the surrounding rock debris or lining temperature, with the unit of °C; μ: the friction coefficient, related to the tool material and the rock type; p: the tool-rock contact pressure, obtained by the feedback of the force sensor or the robotic arm, with the unit of Pa; v: The cutting tool linear velocity, which is equal to the rotational speed ω multiplied by the radius, with the unit of m / s.

4. The intelligent perception method for shield cutter of rotational speed - temperature - wear according to claim 1, wherein Inputting the feature set and the abnormal component into the physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate, including adding constraints to the loss function and adopting the physical constraint loss function to ensure that the data-driven model does not violate the friction-heat transfer mechanism; The formula of the physical constraint loss function is: ； where, N: The total number of training samples; , yi: The predicted and actually measured wear rate of the i-th sample neural network, unit: mm / h; ΔEi pred : Thermal energy change deduced from the predicted wear volume of the i-th sample, unit: J; ΔEi phys : The physical thermal energy change calculated for the i-th sample according to the friction-heat transfer equation, in J; λ: The weight of the constraint term, which is used to balance the fitting accuracy and physical consistency.

5. The intelligent perception method for shield cutter of rotational speed-temperature-wear according to claim 1, wherein The wavelet packet decomposition of the heat flux distribution model for separating the high-frequency heat anomaly component induced by wear to obtain the abnormal component includes: Using the db4 wavelet basis to decompose the heat flux distribution model to the 4th level to extract the high-frequency heat anomaly components in the 3rd and 4th level subbands.

6. The intelligent perception method for shield cutter of rotational speed - temperature - wear according to claim 1, characterized in that, The LSTM-Transformer model includes a bidirectional LSTM encoder and a multi-head self-attention decoder, where the number of attention heads is not less than 8, which is used to capture the long-term and short-term dependencies of the historical sequence; Online incremental calibration adopts the sample replay strategy, where the single weight update ratio does not exceed 5%.

7. The intelligent perception method for shield cutter of rotational speed-temperature-wear according to claim 1, characterized in that In the process of inputting the feature set and the abnormal component into the physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate, the neural network structure is a three-layer fully connected network. The number of input layer nodes is equal to the sum of the dimensions of the feature set and the abnormal component. The middle layer includes an energy conservation constraint term, and the output layer is a single-node prediction of the instantaneous wear rate.

8. The intelligent perception method for shield cutter of rotational speed - temperature - wear according to claim 1, wherein Performing dynamic threshold determination based on P-value control on the prediction sequence to generate the tool risk warning state. The P-value control calculates the distribution parameters of the prediction error based on the sliding window method and dynamically adjusts the warning threshold to the 95% confidence level.

9. The intelligent perception method of the shield cutter for rotational speed-temperature-wear according to claim 1, wherein Using the warning state to adjust the sampling frequencies and alarm levels of the thermocouple and the encoder to obtain the optimized acquisition strategy. The sampling frequencies are divided into three levels according to the warning state: high risk of ≥20 kHz, medium risk of 10 - 20 kHz, and low risk of 5 - 10 kHz for execution.

10. The intelligent perception method for the rotation speed - temperature - wear of the shield cutter according to claim 1, characterized in that Comparing the measured value of the tool wear after actual replacement with the prediction sequence F to update the weights of the neural network and the time series model online to obtain the optimized model. The model update adopts the incremental learning strategy based on sample replay, and only 5% of the weights are updated each time of calibration.

Images